Dear researchers,
I have some sets of data which were fit with an exponential equation: Y = e^(-aX). The exponential equation parameters (a) for the different sets were estimated using nonlinear regression procedures (nls) in R. I would like to compare these parameters to understand if they are statistically different. Can you help me to understand how I should do this?
I thought of linearising my data through ln-transformation but please, consider I have many zero-values.
Thank you very much in advance for your time.
I am not sure whether this completwly answers your question but you can compare nls() models using the anova() function. See,
https://online.stat.psu.edu/stat485/lesson/19/19.9
Dear Sven Greving
Thank you. As far as I know (please, tell me if I am wrong), you can use anova on nls only if the models are nested. Unfortunately I do not have nested models, I have to compare results of the same exponential model applied to different sets of data.
I fear, you are right about the limits of the anova() function. I howevee wonder why you would need to compare unnested models.
Concerning your ln-problem, I can point you towards the bestNormalize package and the included function with the same name. It provides several normalizing functions including also a slightly modified ln function that can deal with zeros by adding a small constant to all values
Thank you for suggesting the bestNormalize package, I will have a look at it.
I want to compare different models built on different datasets to check if I can use one model to describe more than one datasets or I need different models, which is why I am searching for a test to compare nls models or their slopes.
Reading your research question I zhought abou k-fold cross validation. This approach would allow for model construction on parts of original data before testing the model on unseen data.
Cant you just compare the models in terms of goodness of fit?
Here is some reading on k-fold cross validation
https://machinelearningmastery.com/k-fold-cross-validation/
Do not transform the data. That doesn't do any good; see:
Method Non-linear regression -Why you shouldn't take the logarithms...
You could do the regressions with this easy to use program:
www.lerenisplezant.be/fitting.
You will get the parameters and a good estimation of their precision.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series